Fact · Voice privacy and security
A 2024 study found that when a speech model has seen a recording during training, a noisy version of that recording can make the model transcribe its memorized sensitive transcript.
Checked against its source on October 17, 2026
In the source's own words
“They show that when a record has been seen at training time, the model will transcribe the noisy record with its memorized sensitive transcript.”
Noise Masking Attacks and Defenses for Pretrained Speech Models, Matthew Jagielski et al., arXiv, 2024, Abstract.
Open the original- What it measures
- The attack first sends a partly masked version of an utterance and checks whether the model returns the words from its training data. The authors study pretrained speech encoders as well as ASR models.
- What it does not show
- The study uses models built for research. It shows that memorization can be exploited in such models, not how much a commercial dictation product stores.